Stochastic dynamic programming has been extensively used as a method for designing optimal regulation policies for water reservoirs. However, the potential of this method is dramatically reduced by its computational burden, which often forces to introduce strong approximations in the model of the system, especially in the description of the reservoir inflow. In this paper, an approach to partially remedy this problem is proposed and applied to a real world case study. It foresees solving the management problem on-line, using a reduced model of the system and the inflow forecast provided by a dynamic model. By doing so, all the hydrometeorological information that is available in real-time is fully exploited. The model here proposed for the inflow forecasting is a nonlinear, heteroscedastic model that provides both the expected value and the standard deviation of the inflow through dynamic relations. The effectiveness of such model for the purpose of the reservoir regulation is evaluated through simulation and comparison with the results provided by conventional homoscedastic inflow models.